Interindividual variation in human T regulatory cellsAlessandra Ferraroa, Anna Morena D’Alisea, Towfique Rajb,c, Natasha Asinovskia, Roxanne Phillipsd, Ayla Erguna,Joseph M. Replogleb,c, Angelina Bernierd, Lori Laffeld, Barbara E. Strangerb,c,e, Philip L. De Jagerb,c, Diane Mathisa,1,and Christophe Benoista,1

aDivision of Immunology, Department of Microbiology and Immunobiology, Harvard Medical School, Boston, MA 02115; bProgram in TranslationalNeuroPsychiatric Genomics, Department of Neurology, Brigham and Women’s Hospital, and Division of Genetics, Department of Medicine, Brigham andWomen’s Hospital and Harvard Medical School, Boston, MA 02115; cProgram in Medical and Population Genetics, The Broad Institute, Cambridge, MA 02139;dDepartment of Pediatrics/Immunology, Joslin Diabetes Center, Boston, MA 02215; and eSection of Genetic Medicine, Department of Medicine, University ofChicago, Chicago, IL 60637

Contributed by Christophe Benoist, January 27, 2014 (sent for review December 19, 2013)

FOXP3+ regulatory T (Treg) cells enforce immune self-toleranceand homeostasis, and variation in some aspects of Treg functionmay contribute to human autoimmune diseases. Here, we ana-lyzed population-level Treg variability by performing genome-wide expression profiling of CD4+ Treg and conventional CD4+ T(Tconv) cells from 168 donors, healthy or with established type-1diabetes (T1D) or type-2 diabetes (T2D), in relation to genetic andimmunologic screening. There was a range of variability in Tregsignature transcripts, some almost invariant, others more variable,with more extensive variability for genes that control effectorfunction (ENTPD1, FCRL1) than for lineage-specification factors likeFOXP3 or IKZF2. Network analysis of Treg signature genes identi-fied coregulated clusters that respond similarly to genetic and en-vironmental variation in Treg and Tconv cells, denoting qualitativedifferences in otherwise shared regulatory circuits whereas otherclusters are coregulated in Treg, but not Tconv, cells, suggestingTreg-specific regulation of genes like CTLA4 or DUSP4. Dense gen-otyping identified 110 local genetic variants (cis-expression quan-titative trait loci), some of which are specifically active in Treg, butnot Tconv, cells. The Treg signature became sharper with age andwith increasing body-mass index, suggesting a tuning of Tregfunction with repertoire selection and/or chronic inflammation.Some Treg signature transcripts correlated with FOXP3 mRNA and/or protein, suggesting transcriptional or posttranslational regulatoryrelationships. Although no single transcript showed significant asso-ciation to diabetes, overall expression of the Treg signature wassubtly perturbed in T1D, but not T2D, patients.

immunoregulation | suppression

CD4+FOXP3+ regulatory T (Treg) cells are important medi-ators of immune tolerance, prevent overwhelming immune re-

sponses, and regulate extraimmunological functions (1–3). Theirabsence leads to lethal lymphoproliferation and multiorgan auto-immunity in scurfy mice and in patients with immunodysregulationpolyendocrinopathy enteropathy X-linked syndrome.Treg cells differ substantially from conventional CD4+ T cells

(Tconv) with respect to their transcriptomes. In mice, a canoni-cal “Treg signature” of transcripts that are over- or underex-pressed in Tregs relative to Tconv has been well defined (4, 5).This signature encodes proteins ranging from cell-surface mol-ecules (e.g., IL2RA, CTLA4) to transcription factors [e.g., FOXP3or Helios (Ikzf2)] and includes several molecular mediators ofTreg action (6). The Forkhead family transcription factor (TF)FOXP3 is essential for the specification and maintenance of Tregsand plays an important part in determining the Treg signature (4,7–9). However, FOXP3 is not completely necessary for the dif-ferentiation of Treg cells, and some aspects of the Treg signatureare independent of FOXP3 (4, 10–12). A number of other tran-scription factors have been reported to interact with FOXP3 andto promote Treg function (refs. 13 and 14 and refs therein). Inaddition, different Treg subphenotypes are dependent on differ-ential expression of TFs, such as T-bet, Irf4, or PPARγ (15–17). Wehave recently shown that several of these TFs make up, together

with FOXP3, a genetic switch that locks in the Treg phenotype (13).The transcriptome of Treg cells has been less extensively studied inhumans although early studies indicate that several of the moreprominent members of the mouse Treg signature are also differ-entially expressed in human Treg cells (18–20).Dysregulation of Treg cells has been invoked in the de-

terminism of organ-specific autoimmune diseases such as type-1diabetes (T1D) (21). Experimentally, genetic deficiencies thatreduce Treg numbers or some facets of their function result inaccelerated diabetes in mouse models (22–25), and Treg transfercan be protective (22, 26, 27). Whether Treg defects are directlyimplicated in the determinism of autoimmune disease remainsan open question. Some of the T1D susceptibility loci uncoveredby large genome-wide association studies (GWASs) may plau-sibly influence Treg activity (e.g., IL2RA, IL2, CTLA4) (28). Onecan hypothesize that programmed Treg defects render an in-dividual globally more susceptible to unrestrained activity ofautoreactive cells or that Treg deficiencies occurring locally inthe target organ, perhaps in response to environmental or in-fectious triggers, destabilize the local Treg/effector T equilibriumand allow terminal organ damage (29, 30).It is now recognized that the frequency of FOXP3+ Treg cells

in the blood of human T1D patients is comparable with that ofhealthy subjects (30–32), as in NOD mice (29, 33). WhetherTregs from T1D patients are dysfunctional is controversial (30,32, 34–37), and it is possible that only one facet of their activity isaltered (38). Recent results also suggest that, in T1D patients,effector T cells may be refractory to inhibition by Treg cells (39,40) although this point has also been debated (30, 38). However,

Significance

Control of immunologic tolerance and homeostasis rely onregulatory T lymphocytes that express the transcription factorFOXP3. To characterize the interindividual variation of humanTreg cells, we performed a genome-wide expression and ge-notypic analysis of 168 human donors, healthy or affected bytype-1 or type-2 diabetes (T1D, T2D). We identify cis-actinggenetic variants that condition Treg effector but not specifi-cation genes, and gene clusters that suggest Treg-specificregulatory pathways for some key signature genes (CTLA4,DUSP4). We also identify factors that may control FOXP3 mRNAor protein expression, the specification of the Treg signature,and Treg suppressive efficacy. Although no single transcriptcorrelates with diabetes, overall expression of the Treg signa-ture is perturbed in T1D, but not T2D, patients.

Author contributions: A.F., A.M.D., A.B., L.L., B.E.S., P.L.D.J., D.M., and C.B. designed research;A.F., A.M.D., T.R., N.A., R.P., A.E., J.M.R., and A.B. performed research; A.F., A.M.D., T.R., andA.E. analyzed data; and A.F., L.L., B.E.S., P.L.D.J., D.M., and C.B. wrote the paper.

The authors declare no conflict of interest.1To whom correspondence may be addressed. E-mail: cbdm@hms.harvard.edu.

This article contains supporting information online at www.pnas.org/lookup/suppl/doi:10.1073/pnas.1401343111/-/DCSupplemental.

www.pnas.org/cgi/doi/10.1073/pnas.1401343111 PNAS | Published online March 7, 2014 | E1111–E1120

IMMUNOLO

GY

PNASPL

US

Dow

nloa

ded

by g

uest

on

Nov

embe

r 15

, 202

1

there is an intrinsic limitation to addressing this question ex-perimentally in the human system: Only blood cells are readilyaccessible, and the in vitro suppression assay, the only practica-ble tool for functional evaluation, may not be relevant to thecontrol of diabetes in vivo.Human genetic diversity and its adaptation to novel environ-

ments during out-of-Africa migrations have strong impacts ongenes of the immune system (41). Indeed, under the strong se-lective pressure elicited by pathogens, immune system genes arethose that show the strongest marks of adaptation and positiveselection for variant alleles in different populations (42). In-flammatory disease-associated variants such as those underlyingT1D are enriched in signatures of positive selection (43). Some ofthese variants are eQTL (expression quantitative trait loci) thataffect transcriptional rate or mRNA stability (44) and may me-diate the effect of inflammatory-disease susceptibility loci (45).Little is known about the range of genetic and epigenetic

variation in Treg cells within the human species. Various studieshave found a wide (up to fourfold) range of variation in the pro-portion of FOXP3+ Treg cells in healthy individuals (30, 32, 34–37), as is the case in inbred mice (33). Here, we have assessed theinterindividual transcriptomic variability in Treg cells as this in-terindividual variance encompasses and integrates all of the ge-netic, epigenetic, environmental, and stochastic influences thatgovern the Treg transcriptome of an individual. This study wasperformed on cohorts of healthy subjects and patients with estab-lished T1D, aiming to tease out aspects of this variability thatcorrelate with autoimmune diabetes.

ResultsGenetic/Genomic/Immunologic Profiling. To explore the extent ofvariation in Tregs among human individuals, we performeda broad genomic, immunologic, and genetic profiling, with a par-ticular emphasis on T1D. We collected a sample set from 229individuals, comprising 83 T1D patients, 46 type-2 diabetes (T2D)patients, and 100 healthy controls (Dataset S1). Recruiting wassplit into three consecutive cohorts over 3 y because of financiallimitations, which also allowed adjustments in composition andexperimental design. Donors with T1D were sampled well afterdiabetes onset (only one donor <1 y, and 95% >5 y) to avoidconfounders from inflammation and unstable glycemia surround-ing onset. Donors with T2D were included in cohorts 2 and 3 toflag changes in gene expression secondary to hyperglycemia and/orinsulin treatment. Six donors (cohort 2) were sampled on two in-dependent occasions, several weeks apart, to assess the stability ofthe traits observed. Age, sex, body-mass index (BMI), age of T1Donset, disease duration, glycated hemoglobin (HbA1c), and insulinrequirement were recorded. Immunologic profiling included flow-cytometric determination of Treg frequency among CD4+ cells,levels of FOXP3 protein, and the proportion of CD25+ Tregs. Theexpression of Helios and other Treg subset markers (CD39, CCR6,CXCR3, and CCR4), were measured for some donors, as was Tregin vitro suppressive activity.The expression-profiling arm compared freshly isolated Treg

and Tconv cells from each individual. Treg (CD4+CD25hiCD127−)and Tconv (CD4+CD25−CD127+) cells were purified from pe-ripheral blood mononuclear cells (PBMCs) by two rounds of flowsorting (Fig. S1) (sorted Tregs were mainly 80–90% FOXP3+). Toensure a full representation of these CD4+ populations, they wereprofiled without further split. RNAs from these cells were profiledon Affymetrix ST1.0 microarrays, one batch of microarrays cor-responding to each recruitment cohort. Of the profiles determined,we retained 168 pairs with robust Treg and Tconv data (60 T1D, 30T2D, 78 healthy controls). There were interbatch variation be-tween the cohorts, and most of the bioinformatics analyses wereperformed within each cohort and then cross-checked for concor-dance. When the full power of the entire dataset was needed (e.g., tocompare T1D donors and controls), we used a generalized linear

mixed model to fit the data, using the residuals after fit to thebatch variable.

The Treg Cell Signature and Its Variability. The large number ofTreg and Tconv cells profiled allowed a robust definition of thetranscripts that distinguish human Treg and Tconv cells, group-ing transcripts that were differential in the population means orin at least 20% of the individuals. This core Treg signature in-cluded 194 and 192 probes over- and underexpressed in Tregsrelative to Tconv (Treg-Up and Treg-Down, respectively; listedin Dataset S2). However, because of the power provided by 168individuals, significant Treg/Tconv differential expression wasdetected for 7,975 probes [at a false discovery rate (FDR) ≤10−3],indicating that the stronger differences resonate deeply across thegenome. The most differentially expressed genes included the“usual suspects” of Treg biology, FOXP3, CTLA-4, IKZF2, andIL2RA, overexpressed in Tregs, or ID2 and THEMIS overex-pressed in Tconv cells (Fig. 1A). The signature has a high con-cordance (>73%) with a quality Treg signature derived previouslyon a different microarray platform (18). The signature encodesproducts involved in a variety of processes and cellular localizations.We then asked how the expression of the Treg signature genes

differs between individuals. Fig. 1B displays for individual donorsthe Treg/Tconv expression ratio of Treg signature genes, rankedby mean ratio. Several points can be made. First, the signaturewas generally shared among all individuals. Second, some indi-viduals showed marked departure from the population average,with lower or even inverted Treg/Tconv ratios for some genes.Third, there was a range in the variability of individual genes.Some departed little from the population mean, including someof the key defining transcripts (e.g., FOXP3, IKZF2) (Fig. 1C);others were more divergent, denoted by vertical streaks in Fig.1B, exemplified by ENTPD1 (encodes CD39, a Treg effectormolecule) (46, 47). Overall, there was more interindividual var-iability in overall abundance for Treg signature genes than forother transcripts, as indicated by the distribution of coefficientsof variation of signature genes relative to expression-matchedtranscripts that are equivalently expressed in Treg and Tconvcells (Fig. 1D).These results indicate that there is a defined Treg signature

generally shared across the population, but with more inter-individual variability than the transcriptome norm. This variancemay reflect variation of single genes, or in modules and pathwaysthat define the Treg-cell subphenotypes.

Coregulation of Treg Signature Genes. Eukaryotic transcriptomesare organized into clusters of coexpressed genes, responding inconcert to regulatory cues through shared control mechanisms(48). The interindividual variation across these datasets providedan opportunity to identify such clusters among Treg signaturegenes in the two parallel datasets from CD4+ Treg and Tconvcells and to identify connections that are shared in these tworelated cell types, and those that are specific (Fig. 2A). The per-turbations introduced in the regulatory network by genetic, epi-genetic, or environmental variance should result in detectablecorrelation between coregulated gene pairs in Treg cells. Regu-latory circuits that have the same structure but are tuned to dif-ferent levels should yield similar correlated modules in Treg andTconv datasets, but regulatory interactions that do not operate inTconv cells should show up as lower correlation coefficients forthe corresponding gene modules in the Tconv dataset.We first constructed a matrix of gene–gene correlation for all

Treg signature genes and used a partitioning clustering algorithmto group those genes exhibiting multiple correlations into clus-ters, optimizing the cluster structure by taking into account bothTreg and Tconv datasets (Materials and Methods) (similar cor-relations were observed when each cohort was analyzed in-dependently, indicating that the correlations were robust and not

E1112 | www.pnas.org/cgi/doi/10.1073/pnas.1401343111 Ferraro et al.

Dow

nloa

ded

by g

uest

on

Nov

embe

r 15

, 202

1

driven by batch effects). The results (Fig. 2B and listed in DatasetS3) showed an array of gene clusters, which predominantly butnot exclusively grouped either up- or down-signature genes.Several points could be made. First, there wasn’t one dominantcluster grouping FOXP3 and the canonical Treg signature genes.Instead, the latter are dispersed into different clusters (clusters 4,7, 8, and 11 for FOXP3, IL-2RA, CTLA-4, and IKZF2, re-spectively), FOXP3 itself belonging to a small cluster (togetherwith IKZF4, TIGIT, and the proliferation-controlling miRNAprecursor MIR21). These observations suggested that the majorTreg signature transcripts do not vary in lockstep within Tregsand that variations in FOXP3 expression are not the uniquedriver of the Treg signature tone. Second, side-by-side compari-son of the maps of Fig. 2 showed that the patterns of gene–genecorrelation were partially different in Treg and Tconv cells: Someclusters were equally “tight,” with comparable intracluster simi-larity in both cell types (e.g., cluster 2, 3, 5, or 7). Several otherclusters were markedly less connected in the Tconv than in theTreg datasets (e.g., clusters 1, 8, or 12), likely denoting regulatoryconnections that are specific to Treg cells. Third, the cluster map

was very similar for Treg cells from T1D patients and controlindividuals (Fig. S2), indicating that these structures are robust,with comparable relationships in unrelated donors, and that thesedifferences in the regulatory network of Treg and Tconv cells donot relate to T1D.

Genetic Basis for Variability in Treg Signature Genes. We then per-formed an expression quantitative trait locus (eQTL) analysis toidentify and characterize the cis-acting genetic contribution tointerindividual variability in expression levels of Treg signaturegenes. Genotypes at 951,117 SNPs were determined for 65 indi-viduals (Exome and ImmunoChip, providing genome-wide cov-erage with additional weight on exome and loci of immunologicrelevance). Rigorous quality control included genotype call rate,sex misidentification, Hardy–Weinberg Equilibrium testing, andMinor Allele Frequency (MAF) < 0.01; eight individuals and84,461 SNPs were filtered out from later analyses. These genotypeswere used to search for cis-eQTLs, detected as association of RNAexpression levels (residuals of a mixed-model fit to batch, age, anddiagnosis to remove these confounders) with SNPs in a ±1-Mb

500

5000

200

1000

ID2 CCL5THEMIS

500

2000

LRRN3

100

1000

HLA-DRA

100

1000

TregTconv

D

Coefficient of variation

Neutral genesTreg sig genes

0 0.2 0.4 0.6 0.8 1

p=10-48

FOXP3

100

5000

IKZF2

100

1000

ENTPD1 (CD39)

100

1000

C

Treg-down genes

FC m

ean

expr

essi

on T

reg/

Tcon

v (lo

g2)

B Treg-up genesTreg-up genesTreg-down genes

p-va

lue

Mean Treg/Tconv FoldChange

194 Treg-up192 Treg-down

NELL2

THEMIS

TMEM71

DENDD5A

ANK3

CD40L

IL2RACTLA4

IKZF2

FOXP3

RTKN2

IL2RB

ENTPD1

HLADRA

UTS2

A

200.25 1 4

10-24

10-148

10-117

10-66

10-55

0 100 200 400300Genes

0.02

0.1

0.5

50

10

5

20

2

0.2

0.05

FCRL3

TIGIT

Exp

ress

ion

(AU

)

Fig. 1. Interindividual variability in Treg signature transcripts. (A) Gene-expression profiles were generated from purified Treg and Tconv cells from blood of168 individuals, and the average expression values are compared. (B) Treg signature transcripts are ranked according to mean differential expression (reddots), and the Treg/Tconv ratio for each individual of Cohort3 is plotted. (C) Normalized microarray expression levels in Tconv and Treg cells of selected genes (ona linear scale from 20 to 20,000, where 120 represents a 95% probability of expression); each dot represents 1 of 55 donors from cohort 3. (D) Histogram of gene-wise coefficient of variation across the datasets for Treg signature genes or for a random expression-matched set of genes (P value, Wilcoxon rank-sum test).

Ferraro et al. PNAS | Published online March 7, 2014 | E1113

IMMUNOLO

GY

PNASPL

US

Dow

nloa

ded

by g

uest

on

Nov

embe

r 15

, 202

1

window around each gene. Although this study was not poweredfor a high sensitivity, a number of strong eQTLs emerged (Fig. 3Aand listed in Dataset S4): 105 and 110 genes showed at least oneeQTL at 0.001 permutation threshold with a false discovery rate(FDR) of 9% (49) in Treg and Tconv cells, respectively. As ex-pected, many of those transcripts showed several associated SNPs,in linkage disequilibrium and spread across the gene (e.g., ENTPD1,FCRL1, and CD52) (Fig. S3). Interestingly, cis-eQTLs were notdetected for the most characteristic transcription factors, in partic-ular FOXP3 or IKZF2 (Fig. 3B and Dataset S5).Many of the eQTLs detected were shared in Treg and Tconv

cells, as illustrated by the comparison, in Fig. 3B, of the P valuesof the best eQTL for each gene in both cell types. Some eQTLs,however, were predominantly or uniquely significant in Treg or

Tconv cells (predominantly in Tregs for Up signature genes, andvice versa). We queried the Treg and Tconv eQTLs for SNPspreviously associated with immune or inflammatory diseases(Dataset S6). Only a few overlaps were found (e.g., ERAP2 inCrohn disease), and these overlaps concerned only the less sig-nificant GWAS hits.Finally, we asked whether genetic variation accounts for all of

the variance observed in the Treg signature. We computeda Variability Score (VS) for each gene, from the variance ob-served between all individuals relative to the variance observedbetween repeat samples drawn from the same donors (Materialsand Methods). The variance within these repeats from the samedonors incorporates both experimental noise and short-termfluctuations, such that the VS should mainly reflect interindividual

B

A

Genes clusters in Treg

1 2 3 4 5 6 7 8 9 10 131112 Genes clusters in Tconv

1 2 3 4 5 6 7 8 9 10 131112

-1

Treg Tconv

Nodes with comparable connection inboth cell-types yield similar correlations

Cell-specific nodes are identifiable by

different gene-gene correlations

Inter-individual variation“perturbs” the networks

1Cluster1

SPSB1TNFRSF9NUSAP1TRIM16LTRIM16

CCR10CSF2RB

TRIB1DUSP4

Cluster12ATHL1SBF2SOS2ZCCHC14RBMS3

RASGRF2PIONTMEM71

Cluster8TNFSF10CTLA4GBP2SLC16A1F5

SMC4SMC6BARD1CCDC141TFRC

CCNG2HPGD

TOXIL7

STAMARID5B

GLCCI1

HELLSCARD16ST8SIA6PTPLARTKN2

GLCECDHR3

CEP78

WEE1PTPRJ

Clusters :

Clustersimilarities : 0.58 0.250.52 0.490.49 0.460.21 0.130.62 0.610.2 0.120.51 0.480.53 0.330.27 0.250.39 0.310.68 0.52

0.62 0.30.58 0.54

Fig. 2. Coregulated gene clusters in Treg and Tconv cells. (A) Experimental concept. (B) Treg signature genes were tested for correlated expression across alldonors, and the 290 genes with >2 correlated genes (at cc > 0.5) were biclustered within the Treg and Tconv expression datasets. Heatmap representation ofthe gene–gene correlation coefficients, similarly clustered in the Treg and Tconv datasets. The Cluster Similarity score is the mean of Pearson correlationcoefficients between members of each cluster, computed in Treg and Tconv cells. Members of the Treg-specific clusters are shown.

E1114 | www.pnas.org/cgi/doi/10.1073/pnas.1401343111 Ferraro et al.

Dow

nloa

ded

by g

uest

on

Nov

embe

r 15

, 202

1

variation that is stable over time. Plotting this VS against theeQTL score [–log10(P value)] for all genes shows that the geneswith strong eQTLs have a high variability score, providing valida-

tion of this metric. On the other hand, most genes with high VS hadno significant eQTL (Fig. 3C) (only 21.9% genes with VS >0.75had an eQTL with score >4). With the caveat that the limitedpower of this small dataset would miss weak genetic effects, thisdichotomy suggests that the high interindividual variability in Tregsignature genes has a cis-acting genetic basis for some geneswhereas other genetic and or nongenetic factors influence manyother transcripts.

Nongenetic Variation in Treg Signature Transcripts. We then askedwhether the expression of Treg signature genes varied withindividual characteristics of the donors. It has been unclearwhether the proportion of Tregs increases with age in humans, asit does in mice (50). We first analyzed the frequency of Tregs(% CD25+FOXP3+ cells among CD4+ T cells) relative to age, sex,and BMI. None of the traits showed a significant correlation withTreg frequency (Fig. S4 A–C). To study the association of age,sex, and BMI with the expression of Treg signature genes, we fitthe expression of each Treg signature gene in a linear mixedmodel with age, sex, and BMI as explanatory variables, and re-trieved the coefficient for each term (to ensure that diabeteswould not influence the outcome, values for T1D patients andhealthy controls were fit independently). There was a strong biasof the coefficients for the age variable, opposite for Up andDown Treg signature genes, indicating a positive associationbetween age and the expression of Treg signature genes. Thisobservation was true for both T1D patients and control subjects(Fig. 4A). This bias involved most of the Treg signature, althoughmore marked for some clusters than for others (Fig. S4G). Wecalculated “Treg indices” by averaging the normalized expressionof all Treg signature genes. The Treg Up index increased grad-ually, with roughly a 50% increase between 20 and 60 y of age(Fig. 4B). A parallel trend was noted, in both Tconv and Treg cells,for the Treg Down index. There was no bias relative to sex, save fora few ChrX- or Y-encoded genes, but we found an association withBMI, which was more marked in controls than in T1D patients(Fig. 4A).These results suggest that the expression of Treg-specific genes

is modulated by age and metabolic parameters, Treg cells ac-quiring a more distinctive phenotype with age.

Transcripts Controlling Treg Numbers and Function. As in mice, theproportion of blood Treg cells varies markedly between humans,but its determinism is unknown. We thus probed whether Tregsignature genes or other elements of the Treg transcriptome mightdetermine these traits. Across our cohorts, the proportion ofFOXP3+ among CD4+ T cells varied within the usual range (2–8%), with no relation to FOXP3 mean fluorescence intensity(MFI). We correlated Treg frequency values with the expressionof each gene across the Treg datasets (after normalizing for age,sex, diagnosis, and BMI). As illustrated in Fig. 5 A and B andDataset S7, the overall expression of Treg signature genes did notassociate particularly strongly with Treg frequency; rather, Tregfrequency correlated with other transcripts not commonly associ-ated with Treg cells but with cell cycle (e.g., histone genes) oractivation (POL2RE, PKM, NDRG1). In addition, none of thegenotyped SNPs associated significantly with Treg proportions.On the other hand, there was a clear bias in the correlation

coefficients between Treg signature expression and the level ofFOXP3 protein expression, assessed as the MFI in flow-cyto-metric analyses (Fig. 5 C and D and Dataset S7). Most of thecanonical Treg signature genes, with the exception of FCRL3,were biased toward low but positive correlation coefficients (andconversely for Treg-down signature genes; χ2 P= 10−14 and 10−40).Transcripts most correlated with FOXP3 MFI likely include directFOXP3 targets, or conversely encode factors that regulate FOXP3itself. As evidence for the former, genes with highest absolutecorrelation to FOXP3 MFI were also enriched in FOXP3-binding

eQTL p.value in Treg

Var

iabi

lity

Sco

re

51010

0.0

0.2

0.4

0.6

0.8

5

BCAS1

FUT10

WBSCR27

SLC12A7

ENTPD1

CENPK

CSF2RBMCF2L2

HOXB2

FCRL1SESTD1

SHAD5CDK6

IKZF4

XBP1

MBTPS1

TCF7

TRAF1

FUT10

RNASET2CTSW

CD52

FCRL1

BCAS1

ENTPD1

SESTD1

CENPK

WBSCR27CSF2RB

B

0 5 10 15

0

5

10

15

eQTL p.value in Treg

eQTL

p.v

alue

in T

conv

C

0

11 12 13 14 1516 1718

192021

22Chromosome

1 2 3 4 5 6 7 8 9 10

5

10

15

20 Tconv

A

Chromosome1 2 3 4 5 6 7 8 9 10 11 12 13 14 1516 17

1819

2021

22

0

5

10

15

20 Treg

IKZF2

FOXP3

eQTL

p.v

alue

eQTL

p.v

alue

Fig. 3. eQTL analysis in Treg and Tconv gene expression. (A) Genome-wideSNP genotypes from 57 individual belonging to cohorts 2 and 3 were testedfor association to expression of neighboring genes in the Treg and Tconvdatasets. The significance of the association is shown [−log10(P value)]; eachdot is an SNP/gene pair. The dotted line denotes the nominal P value forsignificance. (B) Comparison of maximum eQTL P values for genes in Treg andTconv expression datasets. Red or blue dots and labels are Treg-Up and -Downsignature genes, respectively. (C) eQTL P values are plotted against the Vari-ability Score (measure of true interindividual variability, after correcting forexperimental noise).

Ferraro et al. PNAS | Published online March 7, 2014 | E1115

IMMUNOLO

GY

PNASPL

US

Dow

nloa

ded

by g

uest

on

Nov

embe

r 15

, 202

1

sites (extrapolating from FOXP3 binding to orthologs in mice)(14). FOXP3-correlated transcripts include CCR8, a marker ofTreg cells with higher FOXP3 levels (51), but also genes that donot belong to the Treg signature such as CD101, which has beenassociated with Treg potency (52).We compared (Fig. 5E) how different transcripts related to

FOXP3 mRNA (microarray signal intensity) or protein (MFI). Asimilar bias in the correlation with both metrics was observed formuch of the Treg signature genes (diametrically opposite for Upand Down transcripts), as would be expected for genes whoseexpression is controlled by FOXP3 or that control its transcrip-tion (TRIB1, NUSAP1, WNT7A, PTPN13). Interestingly, othersshowed correlation to FOXP3 protein but not mRNA, suggest-ing that they may influence posttranslational processing ofFOXP3 (CCR8, CD101, CD93).It is not possible to measure human Treg inhibitory activity in

vivo, and the in vitro suppression assay serves as a surrogate testfor Treg function. To ask which transcripts might associate withsuppressive activity, we correlated the abundance of each tran-script in Treg cells with suppressive activity of purified blood Tregcells in 23 donors (Fig. 5F) (for logistic reasons, this analysis wasperformed only on cohort 3 donors). No single gene stood out,perhaps with the exception of TIGIT, already reported as im-portant for suppressive function (53, 54). However, there wasa general skew of Treg Up signature genes, most of which werepositively associated with suppression (Fig. 5G) (Wilcoxon rank-sum test P = 6 × 10−7). In contrast, transcripts underexpressed inTreg cells were not biased with regard to suppression.

Differential Expression of the Treg Signature in T1D Patients andControls. To search for variation in the transcriptomes of Tregor Tconv cells that would track with T1D in our cohorts, we firstperformed a simple marker-trait association, comparing the ex-pression level of individual genes in T1D donors and age-matched healthy controls. As illustrated in Fig. 6A, no individualgene scored at a genome-wide level of significance [−log10(P) ≥5.5], in either Treg or Tconv cells, with the exception of LARP4(a member of the La ribonucleoprotein domain family) andTLR1 in the Tconv transcriptome. Neither did the T1D candidategenes highlighted by large GWAS studies show biased expression(red dots in Fig. 6A).To assess whether T1D might be associated with changes in the

Treg signature as a whole, we compared the population-wide meanand variance between age-matched T1D and healthy controls forall Treg signature genes. On the “volcano plot” displayed in Fig.

6B, no particular gene stood out. However, the signature as a wholewas skewed, with underexpression of a majority of Treg-up signa-ture genes in T1D individuals (P = 3 × 10−5), even if the changeswere very subtle for each individual transcript [mean FoldChange(FC) in the 0.9–0.95 range]. T2D patients compared with age-matched controls did not show this skew (Fig. 6C) (indeed, theopposite). We found no correlation between Treg signature andHbA1c levels, confirming that the skew observed in T1D patientswas not a result of confounding factors such as insulin treatmentor hyperglycemia.

DiscussionGiven the widespread functions of Treg cells, interindividualvariation in Treg numbers and function in humans should havea significant influence on individual resistance to infections,autoimmune disease, or tumors. The degree of Treg activity canhave deleterious effects at either end of the spectrum (highsuppression leading to low autoimmunity but to stunted anti-infectious responses, and vice versa). In addition, Treg action canhave paradoxically favorable effects on anti-infectious responses(55, 56). Thus, selection acting on loci that condition Treg func-tion probably maintains equilibrated Treg function, with wideenough range at the level of the species that diverse challenges canbe met by at least some individuals. The changing environmentalchallenges that have accompanied human migrations have prob-ably molded the gene pool in this regard, as they have immuneresponse genes in general (41, 42). The present study was designedto describe, on a large scale, the extent of Treg functional vari-ability and to track its genetic and transcriptional underpinnings,with a particular emphasis on the association with type-1 diabetes.

Treg Transcriptomic Variability. Differential expression of Tregsignature genes in Treg and Tconv cells was found generallyacross the 168 donors analyzed, none of which had a broad de-viance from the population mean that affected the Treg signa-ture as a whole. On the other hand, significant variance in theabundance of individual Treg signature transcripts was observed,greater than the genome-wide mean. This variance was uneven,more pronounced for Treg genes associated with suppressivefunction than for lineage-specification genes such as FOXP3 orIKZF2 (although these did show some variation, at the level ofmRNA or protein abundance). A particularly clear example ofvariation in a Treg effector transcript is ENTPD1, which encodesCD39, an important ecto-ATPase in the adenosinergic pathwayof suppression. It is not part of any of the Treg-coregulated gene

Age Gender BMIC

oeffi

cien

t in

T1D

A

-0.015 0 0.015

-0.015

0

0.015

-0.02 0 0.02-0.02

0

0.02

Coefficient in controls-0.02 0 0.02

-0.02

0

0.02

Donor age

B Treg Tconv

p=2x10-9

p=7x10-13p=0.02p=NS

p=4x10-25

p=9x10-4

20 30 40 500

1

2p= 0.0094

p= 0.0004

Treg-down index0

1

2p= 0.00031Treg-up index

Treg-up genes Treg-down genes

Fig. 4. Bias in Treg gene expression with age. (A) Gene expression values in the Treg dataset were fitted in a mixed-model with donor age, sex, and BMI asexplanatory variables, independently for T1D and healthy controls, and the coefficients of the fit were plotted (P values were computed from a χ2 test of thedistribution of positive and negative coefficients). (B) A “Treg index” was calculated for each individual (by averaging normalized expression values of allgenes in the Treg-Up or -Down signatures) and plotted against the donor age, for Treg and Tconv cells.

E1116 | www.pnas.org/cgi/doi/10.1073/pnas.1401343111 Ferraro et al.

Dow

nloa

ded

by g

uest

on

Nov

embe

r 15

, 202

1

clusters and varies largely independently, which is consistent withthe strong cis-eQTLs detected at the locus (also previously ob-served in ref. 57). Does the wide range of variation conditionedby this variant reflect neutral genetic noise or differentialrequirements for activity of Treg adenosinergic suppression indifferent infectious or tumor contexts? CD39 has widespreadexpression in, and relevance to, platelet function, coagulation,and nervous-system regulation, such that the evolutionary driveof this variability may not be immunological.From where does this variability arise? Some was due to cis-

acting genetic variants, with highly significant cis-eQTLs emerging,even from a relatively small sample size. Here again, eQTLs werefound in genes involved in effector or modulator functions(ENTPD1, BCAS1, FCRL1) rather than lineage-determining fac-tors. Some of these eQTLs were predominantly manifest in Tregs,particularly those affecting Treg up signature transcripts, eitherbecause these transcripts are expressed only at low levels in Tconvcells, with correspondingly less power for eQTL detection, or be-cause these SNPs directly affect regulatory mechanisms that areuniquely active in Tregs. Many Treg eQTLs were shared with

Tconv cells. With one exception (SESTD1), these shared eQTLsdid not affect differentially expressed signature genes, but corre-spond to functions that are generic to CD4+ T cells (consistentwith this notion, this “sharing” was more pronounced in this studythan when eQTLs in more distantly related cell-types are com-pared) (58).However, an important fraction of the variability does not have

a readily discernable genetic component. Most of the transcriptswith a variability score in the same range as those with strong cis-eQTLs do not show any hint of sizeable cis control. Some of thesesituations might correspond to trans-QTLs (i.e., association oftranscript expression to genetic variants on other chromosomes),which the study was not powered to detect; but trans-eQTLeffects are generally weaker than cis-eQTLs and would be un-likely to account for a large amount of the variability. Thus, wesuggest that much of this variability stems from nongenetic ori-gins: from inheritable epigenetic traits such as parental imprint-ing, from infectious and other immunologic challenges in theindividual’s history, or from other physiological influences. Anexample is provided by the effect of age, with a clear increase in

Treg-up genes Treg-down genesD

0.1 0.5 1 5 10

TRIB1

FOXP3

CTLA4

FDPSL2A

EPHA4UBAC2

MGAT4A

IK2F2

FCRL3

ZNF70

LDLRCCR8CD101

CD93 IL2RA

Cor

rel.

to F

OX

P3

MFI

Mean Treg / Tconv FC

-0.6

0

0.6

-1

-0.5

0

0.5

1C

Cor

rel.

to F

OX

P3

MFI

0.1 0.5 1 5-0.6

0

0.6

10

PKM2

FOXP3

CTLA4

HIST1+12BG

LRRN3

UCP3

POL2RF

IK2F2NELL2

PRDM2

NDRG1

ENTPD1

EPHB6FADD

Cor

rel.

to T

reg

Freq

uenc

y

Mean Treg / Tconv FC

BAC

orre

l. to

Tre

g Fr

eque

ncy

-1

-0.5

0

0.5

1

-0.6 -0.4 -0.2 0 0.2 0.4 0.6

-0.6

-0.4

-0.2

0

0.2

0.4

0.6

p= n.s.

p<10-3

Correlation to FOXP3 Protein

Cor

rela

tion

to F

OX

P3

mR

NA

TRIB1NUSAP1

CD101

CD93

CCR8

BHLHE40PELI2

WNT7APTPN13

E

Gene rank

Cor

rela

tion

to s

uppr

esso

r act

ivity

-0.5

0

0.5

GTreg-up genes Treg-down genes

Mea

n Tr

eg /

Tcon

v FC

-1 -0.5 0 0.5 1

0.5

1

5

RGS1

CR1

FCRL3

F5SELP

TAF5L

ENTPD1IL2RA

ST8SIA6

ANK3

RTKN2

BACE1

METTL7A

NELL2

C14orf181

CCL5

MAPRE2

DSC1

ADAT3LTBP4PGLYRP1

ID2

---

CTLA4

ATICNEURL3

NCRNA00152

SLC40A1

IKZF2

CCR4

MLH1

TIGIT

RHOH

THEMIS

LOC441204

LRRN3

TRBV23OR9-2ZNF673

TAZ

FOXP3

HLA-DRA

ZBTB22

F

Correlation to suppressor activity

Genes

Genes

Fig. 5. Relationships to FOXP3 expression and Treg function. (A) Expression of individual genes was correlated to Treg frequency (proportion ofFOXP3+CD25+ cells within CD4+ cells), and the correlation coefficients were plotted (blue and red dots, Treg-Up and -Down signature genes, respectively). (B)Plot of the coefficients against the Treg/Tconv expression ratio. (C and D) As in A and B, but vs. FOXP3 mean fluorescence intensity in CD4+CD25+CD127lo cells.(E) Correlation of all genes to FOXP3 mRNA expression in the Treg dataset vs. correlation to FOXP3 protein [from flow cytometry MFI; blue and red dots, Treg-Up and -Down signature genes, gated on FoldChange > 2)]. (F) Correlation of gene expression vs. Treg suppressor activity, measured in a subset of 22 donors.(G) Ranked plot of correlation coefficients for all genes (black dots) or for the Treg-Up and -Down signature genes (red, blue).

Ferraro et al. PNAS | Published online March 7, 2014 | E1117

IMMUNOLO

GY

PNASPL

US

Dow

nloa

ded

by g

uest

on

Nov

embe

r 15

, 202

1

the overall expression of Treg signature transcripts in older indi-viduals. In essence, Treg cells become more Treg-like with age.This change may correspond to a cell-autonomous evolution ofthe Treg population with time, perhaps as a gradual selection ofclones with better survival characteristics, perhaps an elementof “immuno-senescence.”Alternatively, it may be a reaction to thechronic inflammatory tone (with higher levels of proinflammatorycytokines and adhesion molecules) found in aging individuals (59).Some of this variability could be grouped into coregulated sets

of genes. Most of the coregulation patterns were similar in Tregand Tconv cells. This result is not unexpected because these arevery closely related differentiated cell types, but it gives confi-dence as to the robustness of these structures. Further, thesemaintained patterns of coregulation indicate that, even thoughthese Treg signature genes are differentially expressed in Treg andTconv cells, their regulatory connections are similar; differencesin expression may simply reflect a different tuning of thesepathways in the two cell types. On the other hand, some of thepairwise gene–gene correlations or coregulated clusters wereunique or at least highly preferential to the Treg dataset, andthese may reflect Treg-specific circuits. These Treg-specific rela-tionships involve some key Treg-specific genes such as CTLA4,TNFRSF9 (4.1BB), or DUSP4.Strong interindividual variation in the proportion of Treg cells

among blood CD4+ T cells has been observed in many studies andwas also seen here (1.9–8.9%). Perhaps counterintuitively, therewas no association between this frequency and the intensity of theTreg signature genes in general, nor with FOXP3 expression inparticular. This conclusion was confirmed by the lack of correla-tion between Treg frequency and FOXP3 MFI. Stronger Tregs donot make for more Tregs. Some transcripts that did correlate withTreg frequency are plausibly related to cell division or survival, butothers such as FADD or PDRM2 may denote more specific effectsand will be interesting to track in replicative studies.In contrast, there was a clearly biased association of Treg sig-

nature genes with the intensity of FOXP3 protein expression.

Network analysis (Fig. 2) showed that FOXP3 is not the uniqueand dominant driver of the network, in keeping with the nowaccepted notion that FOXP3 is a key important player but not theunique player in Treg signature specification. Some of this asso-ciation likely reflects the importance of FOXP3 in controlling theexpression of Treg signature genes, and many of these (e.g.,CTLA4, TRIB1, RTKN2, IKZF2) are direct FOXP3 targets inmouse cells (14). Others may actually be direct regulators ofFOXP3 transcription or of protein stability. In this regard, it isinteresting to note that there is more interindividual variability inFOXP3 protein than in FOXP3 mRNA and only a minor re-lationship between the two, suggesting translational or posttransla-tional control. Some of these potential modulators of FOXP3 in-clude genes such as CCR8 or CD101, which do not belong to theTreg signature but have been previously tied to Foxp3 expressionand/or Treg potency (51, 52), and several genes involved with lipiduptake and metabolism (LDLR, TRIB1), possibly reflecting a rolein cellular lipids in controlling FOXP3, or genes of yet unrecognizedrelevance that may warrant further exploration (CD93, MGAT4A).Interestingly, CCR8 or CD101 seem to correlate with FOXP3protein but not mRNA, possibly suggesting that they are involvedwith FOXP3 posttranslational regulation.

Treg Genomic Variability and Disease. Given their importance inimmunoregulation, the notion that Treg defects have a causalimplication in the susceptibility to autoimmune diseases hasattracted much interest, with arguments based on the relevanceof some GWAS hits to Treg physiology (IL2RA, CTLA-4) and onreports, albeit not always consistent, of defective Tregs in someautoimmune diseases such as multiple sclerosis. Do the param-eters of Treg variability analyzed here relate to T1D?At the level of individual genes, there was no such connection.

None of the transcripts individually showed significant associa-tion with T1D in the Treg or Tconv datasets, even at relaxedthresholds. This result is probably not surprising, as one mightnot expect that any one transcript would stand out, given the

68126

T1D Treg vs HC

P-va

lue

Mean FoldChange

P=3.1x10-5

14450

P=1.5x10-11

126 66

112 80

Treg-up genes Treg-down genes

P=0.02

P=1.5x10-5

10-4

1

10-1

10-2

10-3

P-va

lue

Mean FoldChange

110-4

1

10-1

10-2

10-3

A B

C T2D Treg vs HC

10.7 1.5

10.7 1.510.7 1.5

0 2 4 6

0

2

4

6

LARP4

ARHGAP9

RAB11FIP4

TLR1

MYC

FAM84B

PLCXD1

Treg T1D vs Ctl P-value (-log10)

Tcon

v T1

D v

s C

tl P

-val

ue (-

log1

0)

RASGRP1

FOXP3

CTLA4

IL27

0.7 1.5

Fig. 6. Underexpression of Treg signature genes in Treg cells from T1D patients. (A) Expression values of each gene in the Treg and Tconv datasets werecompared in Tregs from T1D patients or healthy controls. The nominal threshold for corrected genome-wide significance is shown as a dotted line. (B) MeanFold Change vs. t test P value for the comparison between T1D and healthy donors for Treg-Up (Left) and -Down (Right) signature genes. (C) As in B,comparing T2D and healthy donors.

E1118 | www.pnas.org/cgi/doi/10.1073/pnas.1401343111 Ferraro et al.

Dow

nloa

ded

by g

uest

on

Nov

embe

r 15

, 202

1

cellular and regulatory complexity of T1D pathogenesis. Simi-larly, the gene-coregulation patterns of Fig. 2 were largely su-perimposable for T1D and control donors. On the other hand,there were clear indications that combinations of genes are af-fected in the comparison of T1D patients to age-matched controls:When the Treg signature was tested as a whole, deviations in in-dividual genes that were individually very small showed a signifi-cant collective bias.We cannot formally rule out that some of these combinatorial

signals are a secondary consequence of the inflammation andglycemic dysregulation that result from T1D. On the other hand,several arguments argue against this interpretation: (i) patientsin this study had established and stable disease, several or manyyears removed from the inflammatory perturbations that ac-company onset; (ii) T2D patients, half of which were insulin-dependent, were included as controls and did not show theseeffects; (iii) there was no relationship between the signature andglycemic control as reflected by HbA1c; and (iv) the decreasedexpression of Treg signature genes in T1D patients was preciselythe opposite of that seen with both aging and T2D, both of whichhave a marked inflammatory component. This bias could not beascribed to any cis- or trans-acting genetic locus. More generally, wesuggest that the “state” of Treg population, as reflected by itsmRNA expression profile, incorporates some genetic elements butalso many nongenetic ones: parentally transmitted epigenetic traits,epigenetic remodeling during development, microbiome influence,

and immunologic history. It is this integration that may define thesusceptibility of a given individual to autoimmune disease.

Materials and MethodsDetailed methods are provided as SI Materials and Methods.

Briefly, blood samples were obtained from 229 individuals (83 T1Dpatients, 46 T2D patients, and 100 age-matched controls). PBMCs wereprepared and double-sorted immediately using a strict standard operatingprotocol to minimize environmental, technical, and circadian variation, intoTreg (CD4+CD25hiCD127lo) and Tconv (CD4+CD25−CD127hi) pools. Trizol-prepared RNAs were profiled on Affymetrix ST1.0 HuGene arrays. Data fromthree different cohorts were normalized by Robust Multichip Average andanalyzed in GenePattern Multiplot. Batch effects were removed by firstmeans-normalizing the data of each batch, then using residuals of a general-ized linear model fit to the batch variable. Correlation to donor variables,immunophenotypes, or Treg suppression results were tested in R or S-Plus. SNPgenotyping was performed for 65 donors on the Illumina Infinium Human-OmniExpressExome (951,117 SNPs) and quality-controlled data were analyzedusing PLINK v1.07 software (60). In vitro Treg suppression of anti-CD3 activatedTconv cells was as described (30), using as responder cells frozen aliquots froma single donor.

ACKNOWLEDGMENTS. We thank Dr. D. Koller for helpful discussion; KatieRothamel, Joyce LaVecchio, and Girijesh Buruzula for help with flow cytometryand RNA preparations; and Jeff Ericson, Scott Davis, and Henry Paik for helpwith bioinformatics analyses. This work was supported by Juvenile DiabetesResearch Foundation Grant 4-2007-1057 (to D.M., C.B., and L.L.) and NationalInstitutes of Health Grant RC2 GM09308 (to D.M. and C.B.).

1. Sakaguchi S, Yamaguchi T, Nomura T, Ono M (2008) Regulatory T cells and immunetolerance. Cell 133(5):775–787.

2. Curotto de Lafaille MA, Lafaille JJ (2009) Natural and adaptive foxp3+ regulatoryT cells: More of the same or a division of labor? Immunity 30(5):626–635.

3. Benoist C, Mathis D (2012) Treg cells, life history, and diversity. Cold Spring HarbPerspect Biol 4(9):a007021.

4. Hill JA, et al. (2007) Foxp3 transcription-factor-dependent and -independent regula-tion of the regulatory T cell transcriptional signature. Immunity 27(5):786–800.

5. Feuerer M, Hill JA, Mathis D, Benoist C (2009) Foxp3+ regulatory T cells: differentia-tion, specification, subphenotypes. Nat Immunol 10(7):689–695.

6. Vignali DA, Collison LW, Workman CJ (2008) How regulatory T cells work. Nat RevImmunol 8(7):523–532.

7. Ziegler SF (2006) FOXP3: Of mice and men. Annu Rev Immunol 24:209–226.8. Hori S, Nomura T, Sakaguchi S (2003) Control of regulatory T cell development by the

transcription factor Foxp3. Science 299(5609):1057–1061.9. Fontenot JD, et al. (2005) Regulatory T cell lineage specification by the forkhead

transcription factor foxp3. Immunity 22(3):329–341.10. Gavin MA, et al. (2007) Foxp3-dependent programme of regulatory T-cell differen-

tiation. Nature 445(7129):771–775.11. Lin W, et al. (2007) Regulatory T cell development in the absence of functional Foxp3.

Nat Immunol 8(4):359–368.12. Sugimoto N, et al. (2006) Foxp3-dependent and -independent molecules specific for

CD25+CD4+ natural regulatory T cells revealed by DNA microarray analysis. Int Im-munol 18(8):1197–1209.

13. Fu W, et al. (2012) A multiply redundant genetic switch ‘locks in’ the transcriptionalsignature of regulatory T cells. Nat Immunol 13(10):972–980.

14. Samstein RM, et al. (2012) Foxp3 exploits a pre-existent enhancer landscape forregulatory T cell lineage specification. Cell 151(1):153–166.

15. Zheng Y, et al. (2009) Regulatory T-cell suppressor program co-opts transcriptionfactor IRF4 to control T(H)2 responses. Nature 458(7236):351–356.

16. Koch MA, et al. (2009) The transcription factor T-bet controls regulatory T cell ho-meostasis and function during type 1 inflammation. Nat Immunol 10(6):595–602.

17. Cipolletta D, et al. (2012) PPAR-γ is a major driver of the accumulation and phenotypeof adipose tissue Treg cells. Nature 486(7404):549–553.

18. Pfoertner S, et al. (2006) Signatures of human regulatory T cells: An encounter withold friends and new players. Genome Biol 7(7):R54.

19. Jeron A, et al. (2009) Frequency and gene expression profile of regulatory T cells inrenal cell carcinoma. Tumour Biol 30(3):160–170.

20. Birzele F, et al. (2011) Next-generation insights into regulatory T cells: Expressionprofiling and FoxP3 occupancy in Human. Nucleic Acids Res 39(18):7946–7960.

21. Buckner JH (2010) Mechanisms of impaired regulation by CD4(+)CD25(+)FOXP3(+)regulatory T cells in human autoimmune diseases. Nat Rev Immunol 10(12):849–859.

22. Salomon B, et al. (2000) B7/CD28 costimulation is essential for the homeostasis of theCD4+CD25+ immunoregulatory T cells that control autoimmune diabetes. Immunity12(4):431–440.

23. Chen Z, Herman AE, Matos M, Mathis D, Benoist C (2005) Where CD4+CD25+ T regcells impinge on autoimmune diabetes. J Exp Med 202(10):1387–1397.

24. Bettini ML, et al. (2012) Loss of epigenetic modification driven by the Foxp3 tran-scription factor leads to regulatory T cell insufficiency. Immunity 36(5):717–730.

25. Darce J, et al. (2012) An N-terminal mutation of the Foxp3 transcription factor alle-viates arthritis but exacerbates diabetes. Immunity 36(5):731–741.

26. Tarbell KV, Yamazaki S, Olson K, Toy P, Steinman RM (2004) CD25+ CD4+ T cells, ex-panded with dendritic cells presenting a single autoantigenic peptide, suppress auto-immune diabetes. J Exp Med 199(11):1467–1477.

27. Tang Q, et al. (2004) In vitro-expanded antigen-specific regulatory T cells suppressautoimmune diabetes. J Exp Med 199(11):1455–1465.

28. Pociot F, et al. (2010) Genetics of type 1 diabetes: What’s next? Diabetes 59(7):1561–1571.29. Tang Q, et al. (2008) Central role of defective interleukin-2 production in the trig-

gering of islet autoimmune destruction. Immunity 28(5):687–697.30. Ferraro A, et al. (2011) Expansion of Th17 cells and functional defects in T regulatory

cells are key features of the pancreatic lymph nodes in patients with type 1 diabetes.Diabetes 60(11):2903–2913.

31. Bacchetta R, et al. (2006) Defective regulatory and effector T cell functions in patientswith FOXP3 mutations. J Clin Invest 116(6):1713–1722.

32. Lindley S, et al. (2005) Defective suppressor function in CD4(+)CD25(+) T-cells frompatients with type 1 diabetes. Diabetes 54(1):92–99.

33. Feuerer M, et al. (2007) Enhanced thymic selection of FoxP3+ regulatory T cells inthe NOD mouse model of autoimmune diabetes. Proc Natl Acad Sci USA 104(46):18181–18186.

34. Brusko TM, Wasserfall CH, Clare-Salzler MJ, Schatz DA, Atkinson MA (2005) Func-tional defects and the influence of age on the frequency of CD4+ CD25+ T-cells intype 1 diabetes. Diabetes 54(5):1407–1414.

35. Putnam AL, Vendrame F, Dotta F, Gottlieb PA (2005) CD4+CD25high regulatory T cellsin human autoimmune diabetes. J Autoimmun 24(1):55–62.

36. Long SA, et al. (2010) Defects in IL-2R signaling contribute to diminished maintenanceof FOXP3 expression in CD4(+)CD25(+) regulatory T-cells of type 1 diabetic subjects.Diabetes 59(2):407–415.

37. McClymont SA, et al. (2011) Plasticity of human regulatory T cells in healthy subjectsand patients with type 1 diabetes. J Immunol 186(7):3918–3926.

38. Hughson A, et al. (2011) Uncoupling of proliferation and cytokines from suppressionwithin the CD4+CD25+Foxp3+ T-cell compartment in the 1st year of human type 1 di-abetes. Diabetes 60(8):2125–2133.

39. Schneider A, et al. (2008) The effector T cells of diabetic subjects are resistant toregulation via CD4+ FOXP3+ regulatory T cells. J Immunol 181(10):7350–7355.

40. Lawson JM, et al. (2008) Increased resistance to CD4+CD25hi regulatory T cell-mediatedsuppression in patients with type 1 diabetes. Clin Exp Immunol 154(3):353–359.

41. Barreiro LB, Quintana-Murci L (2010) From evolutionary genetics to human immu-nology: How selection shapes host defence genes. Nat Rev Genet 11(1):17–30.

42. Clark AG, et al. (2003) Inferring nonneutral evolution from human-chimp-mouse or-thologous gene trios. Science 302(5652):1960–1963.

43. Raj T, et al. (2013) Common risk alleles for inflammatory diseases are targets of recentpositive selection. Am J Hum Genet 92(4):517–529.

44. Pai AA, et al. (2012) The contribution of RNA decay quantitative trait loci to inter-individual variation in steady-state gene expression levels. PLoS Genet 8(10):e1003000.

45. Stranger BE, De Jager PL (2012) Coordinating GWAS results with gene expression ina systems immunologic paradigm in autoimmunity. Curr Opin Immunol 24(5):544–551.

46. Deaglio S, et al. (2007) Adenosine generation catalyzed by CD39 and CD73 expressedon regulatory T cells mediates immune suppression. J Exp Med 204(6):1257–1265.

Ferraro et al. PNAS | Published online March 7, 2014 | E1119

IMMUNOLO

GY

PNASPL

US

Dow

nloa

ded

by g

uest

on

Nov

embe

r 15

, 202

1

47. Borsellino G, et al. (2007) Expression of ectonucleotidase CD39 by Foxp3+ Treg cells: Hy-drolysis of extracellular ATP and immune suppression. Blood 110(4):1225–1232.

48. Segal E, et al. (2003) Module networks: identifying regulatory modules and their condition-specific regulators from gene expression data. Nat Genet 34(2):166–176.

49. Churchill GA, Doerge RW (1994) Empirical threshold values for quantitative trait map-ping. Genetics 138(3):963–971.

50. Raynor J, Lages CS, Shehata H, Hildeman DA, Chougnet CA (2012) Homeostasis andfunction of regulatory T cells in aging. Curr Opin Immunol 24(4):482–487.

51. Soler D, et al. (2006) CCR8 expression identifies CD4 memory T cells enriched forFOXP3+ regulatory and Th2 effector lymphocytes. J Immunol 177(10):6940–6951.

52. Fernandez I, et al. (2007) CD101 surface expression discriminates potency amongmurine FoxP3+ regulatory T cells. J Immunol 179(5):2808–2814.

53. Yu X, et al. (2009) The surface protein TIGIT suppresses T cell activation by promoting thegeneration of mature immunoregulatory dendritic cells. Nat Immunol 10(1):48–57.

54. Stanietsky N, et al. (2009) The interaction of TIGIT with PVR and PVRL2 inhibits humanNK cell cytotoxicity. Proc Natl Acad Sci USA 106(42):17858–17863.

55. Lund JM, Hsing L, Pham TT, Rudensky AY (2008) Coordination of early protectiveimmunity to viral infection by regulatory T cells. Science 320(5880):1220–1224.

56. Pace L, et al. (2012) Regulatory T cells increase the avidity of primary CD8+ T cellresponses and promote memory. Science 338(6106):532–536.

57. Stranger BE, et al. (2007) Population genomics of human gene expression. Nat Genet39(10):1217–1224.

58. Dimas AS, et al. (2009) Common regulatory variation impacts gene expression ina cell type-dependent manner. Science 325(5945):1246–1250.

59. Cavanagh MM, Weyand CM, Goronzy JJ (2012) Chronic inflammation and aging: DNAdamage tips the balance. Curr Opin Immunol 24(4):488–493.

60. Purcell S, et al. (2007) PLINK: A tool set for whole-genome association and pop-ulation-based linkage analyses. Am J Hum Genet 81(3):559–575.

E1120 | www.pnas.org/cgi/doi/10.1073/pnas.1401343111 Ferraro et al.

Dow

nloa

ded

by g

uest

on

Nov

embe

r 15

, 202

1